AI Weekly Trends – Highly Opinionated Signals from the Week [W24] 🚀

Jun 23, 2025

This week brought transformative insights from Andrej Karpathy's keynote on Software 3.0, revealing how AI is fundamentally reshaping software development paradigms. While Meta's aggressive talent acquisition strategy creates ripples across Silicon Valley, I'm holding off on discussing the evolving OpenAI-Microsoft tensions until their strategies become clearer. For those interested in deeper analysis on this topic, We explored some hypotheses in this sunday’s "Risorse Artificiali" podcast (Italian only) on Youtube and Spotify.

The rapid evolution of agent-based development, pricing model innovations, and multimodal advances for robotics paint a picture of an industry accelerating toward more autonomous, yet human-collaborative systems.

Trend 1: Software 3.0: Beyond Vibe Coding and Agents

Andrej Karpathy's keynote on how AI is changing software delivered a masterclass that every software developer should absorb. His framework elegantly categorizes our evolution: Software 1.0 (traditional programming), Software 2.0 (neural networks), and now Software 3.0 (where LLMs function as cloud-based operating systems programmable through natural language, what he brilliantly terms "vibe coding.") Rather than pursuing fully autonomous AI agents, Karpathy advocates for "autonomy sliders", a concept beautifully exemplified in tools like Cursor that compensate for AI limitations through human supervision.

The power of partially autonomous applications struck me particularly hard. Consider why a coding agent integrated into an IDE outperforms completely unattended or CLI agents: while it might be less efficient on simple tasks, it excels at complex challenges. Claude Code's magic lies in its iterative approach - exploring the entire solution space until finding what works. This fluid division between machine code generation and human review, planning, and correction creates a development experience that amplifies both human and machine strengths.

Cursor's new $200 Ultra plan reflects this reality. Instead of limiting API calls, they've chosen to limit speed: Pro now offers infinite calls at lower velocity compared to Ultra. This pricing innovation, enabled by long-term agreements with model providers, represents a significant step in making AI development tools more accessible. The economics are clear: AI APIs cost money, and companies need sustainable revenue models.

The concept of "building for agents" from Karpathy confirms a growing sentiment across the industry: we're transitioning from an internet of people, through an internet of things, to increasingly an internet of agents. These aren't mutually exclusive, they must be thoughtfully integrated.

OpenAI's practical guide to building agents emphasizes starting with single agents before multi-agent systems, using manager patterns where one agent coordinates others through tool calls or decentralized handoffs. Key insights include implementing layered guardrails (LLM-based classifiers, regex filters, moderation APIs), designing tools for messy long-horizon tasks, and building human-in-the-loop mechanisms triggered by failure thresholds or high-risk actions.

Security remains fundamental in distributed systems, whether involving intelligent agents or not. The recent changes to the Model Context Protocol, removing JSON-RPC batching support, adding structured tool output support, and clarifying security considerations, underscore this priority. As we build this agent internet, we must ensure robust foundations.

Kimi-Dev represents another fascinating specialization trend, a powerful open-source language model specifically trained for issue resolution. By focusing models on error correction and debugging, we're seeing the emergence of specialized AI tools that excel in narrow but critical domains. This specialization, combined with the iterative approaches Karpathy describes, points toward a future where different models handle different aspects of the development lifecycle.

Key Takeaways for AI Engineers

Software 3.0 paradigm: Natural language becomes the primary programming interface, with LLMs as cloud-based operating systems
Partial autonomy wins: Human-AI collaboration through rich UIs outperforms fully autonomous systems for complex tasks
Agent-first infrastructure: The internet is evolving to accommodate agents as first-class citizens alongside humans and IoT devices
Action Items:
- Experiment with autonomy sliders in your IDE workflows
- Study OpenAI's agent building guide for production implementations

Trend 2: Models: Innovative Architectures and Techniques for Continuous Improvement

Perfectly aligning with Karpathy's Software 3.0 vision, the research community is pushing hard on models that can develop, adapt, and evolve other models. The current limitations (static models, expensive and lengthy training, complicated fine-tuning) are being systematically addressed. Self-adapting language models represent a breakthrough response to this stasis, enabling LLMs to generate "self-modifications" that produce persistent weight updates through supervised fine-tuning. Despite outperforming GPT-4.1 with a smaller model, the approach suffered from catastrophic forgetting and required 15x more tokens than standard inference, but it points toward a future where models autonomously improve through self-generated training material rather than relying on external human-generated text.

Text-to-LoRA from Sakana AI tackles fine-tuning complexity from another angle. Their T2L system can instantly personalize large language models using just a text description: no training data or lengthy fine-tuning required. By compressing hundreds of LoRA adapters into a single network that generates new personalizations on demand, they're democratizing model customization. Making models more dynamic favors adoption, while simplifying fine-tuning makes small models with accessible enterprise costs more viable.

Efficiency drives down inference costs, the major expense for model users and self-hosters. Google's Gemini 2.5 family update exemplifies this trend. Gemini 2.5 Pro has left preview while Flash-Lite enters preview, handling high-volume AI workloads without significant costs. These customized versions now power AI overviews and AI mode in Google Search. While we're still talking about large cloud-executed models, the performance-per-dollar improvements are substantial.

MiniMax's M1 release pushes efficiency boundaries further. Their 456B parameter model uses a hybrid mixture-of-experts architecture with "lightning attention" processing 1 million token contexts (8x DeepSeek R1) while using 25% fewer FLOPs for 100K token generation lengths. Similarly, MiniMax's Hailuo 02 video model features Noise-aware Compute Redistribution architecture, improving training and inference efficiency by 2.5x while tripling parameters and quadrupling training data compared to its predecessor.

Sam Altman has raised the stakes by discussing GPT-5's timeline in recent interviews. Early testers describe it as "materially better" than GPT-4, with enhanced reasoning and agentic capabilities arriving "probably during this summer." Altman also discussed ChatGPT advertising possibilities, drawing a clear line against letting payments influence responses, suggesting ads might appear outside the model's output stream. Another step toward AGI? The convergence of more efficient architectures, self-improving capabilities, and enhanced reasoning suggests we're approaching a qualitative shift in AI capabilities.

Key Takeaways for AI Engineers

Dynamic models emerging: Self-adapting and instantly customizable models address current static limitations
Efficiency breakthrough: New architectures deliver superior performance with significantly lower computational costs
GPT-5 approaching: Summer release promises enhanced reasoning and agentic capabilities
Action Items:
- Explore Text-to-LoRA for rapid model customization
- Benchmark Flash-Lite for high-volume, cost-sensitive workloads

Trend 3: Money and Strategy

Meta's aggressive AGI team formation continues dominating strategic headlines this week, with Sam Altman revealing that Meta reportedly offered OpenAI and Google DeepMind employees compensation packages worth over $100 million. These efforts have been largely unsuccessful, with Altman suggesting that Meta's focus on high compensation rather than the mission of delivering AGI won't create great culture. The talent war escalated further when Meta attempted to acquire Ilya Sutskever's Safe Superintelligence startup, valued at $32 billion. After being rebuffed, Zuckerberg successfully recruited SSI's CEO Daniel Gross and former GitHub CEO Nat Friedman to work on products under Alexandr Wang.

The ripple effects extend beyond hiring. OpenAI is cutting ties with Scale AI following Meta's multi-billion investment in the data labeling startup that brought CEO Alexandr Wang onto Meta's team. Google reportedly plans to follow suit, abandoning Scale as a provider. These moves reveal how talent acquisition strategies are reshaping industry partnerships and data supply chains.

Vibe coding proves itself far from a nerd toy, attracting serious investment from big tech and funds. Google's Release Notes podcast featuring Connie Fan and Danny Tarlow discussing Gemini's coding capabilities shows Google focusing significant efforts on Gemini specifically for vibe coding. The billion-dollar valuations tell the story: Anysphere (Cursor's developer) fields VC offers at over $18 billion, potentially doubling its valuation as the company continues rapid growth in the AI development tools market. Remember that main competitor Windsurf was recently acquired by OpenAI at a $3 billion valuation. These valuations reflect a fundamental shift in how we build software, with AI-augmented development becoming the new standard rather than an experimental approach.

Key Takeaways for AI Engineers

Talent wars intensify: $100M+ offers signal desperate competition for AI expertise
Partnership disruption: Strategic acquisitions force companies to reconsider data providers
Vibe coding mainstream: Billion-dollar valuations confirm AI-augmented development as industry standard
Action Items:
- Monitor partnership shifts for API availability changes
- Evaluate AI coding tools for team productivity gains

Trend 4: Multimodal Models are Functional for Robots' Development

The recent push in developing realistic video generation models with accurate physics simulation represents a critical node for robotics development. World models serve as the foundation for robot spatial movement, while 3D environment generation (closely related to video generation) with realistic physics creates accelerated training grounds for robot movement through reinforcement learning. Meta's V-JEPA 2 announcement exemplifies this approach, particularly excelling at neural perception versus world models for autonomous driving. The key distinction: learning actions from sensor/camera data (behavior cloning) versus constantly building and updating a world representation to understand evolution from the current point and decide appropriate actions. Each approach has trade-offs: the first is scalable but potentially less robust, while the second truly "understands" circumstances, adapting better to novel situations lacking training data.

Odyssey's interactive video showcases another world model application: a hybrid between generated video and videogame where each frame depends on current state and user actions, conditioned by the model's video training rather than game physics code. With 40ms per frame generation and streaming, it potentially represents a new medium entirely.

Google's Veo collaboration on "ANCESTRA" demonstrates practical applications, blending live-action with AI-generated video through customized and motion-matched generative content. Midjourney's V1 launch at $10/month transforms images into video clips with camera movement and animation style controls, significantly cheaper than competitors and capable of generating styles difficult to achieve with other tools.

Research continues optimizing frame-specific chunking for more efficient inference. Physical Intelligence's real-time action chunking addresses a critical robotics challenge: unlike chatbots or image generators, robots must operate in real-time where input-output delays tangibly affect performance. Their algorithm enables seamless real-time execution on any diffusion or flow-based vision-language-action (VLA) model without training changes.

Meanwhile, autonomous vehicles, our most realistic street-ready robots, continue expanding. Waymo plans returning to New York City next month for mapping and testing (human-driven due to state law requirements). With over 250,000 autonomous rides weekly in other cities, Waymo's popularity surge demonstrates real-world multimodal AI deployment at scale.

Key Takeaways for AI Engineers

World models critical: Physics-accurate video generation enables accelerated robotics training
Real-time constraints: Robotics demands new algorithms for seamless, low-latency inference
Convergence accelerating: Video generation, world models, and robotics share fundamental technologies
Action Items:
- Have a look V-JEPA 2 or Odissey for embodied AI applications
- Experiment with Veo…and have fun 🙂

Artificial Code

Discussion about this post